Anthropic Applied AI recommends one-agent loop
Drawing on insights from Anthropic’s Applied AI team, the author argues that replacing multi-agent subagent architectures with a single agent loop using skills and tools often improves latency, cost, and quality. While subagents can still be useful for parallel work or strict context isolation, consolidating them avoids lossy handoffs, subagent sprawl, and conflicting system prompts. The article outlines a migration strategy using evals, tool conversion, and careful merging. Open article →